Dynamic infrared scanning thermography based on CNN: a novel large-scale honeycomb defect detection and classification technique

This paper introduces a highly efficient technique, namely, dynamic infrared scanning thermography (DIST), for detecting defects in large-sized carbon fiber-reinforced polymer/aluminum (CFRP/Al) honeycomb composites. The corresponding test specimen with a high aspect ratio was fabricated for experimental validation by using a DIST system. The pseudo-static matrix reconstruction (PSMR) method and static image sequence processing algorithms were, respectively, employed to pre-process and post-process the experimental data. The results indicate that the DIST method can continuously and effectively detect defects in large-sized CFRP/Al specimens. The respective infrared image dataset was produced, and different convolutional neural network (CNN) models and optimizers were combined for training and comparatively performing automatic defect classification. The obtained results indicate that the combination of the SqueezeNet approach and stochastic gradient descent with momentum (SGDM) is the best when considering the training time as a figure of merit. Such combination provided the accuracy of 99.86% with the time cost of 8.6 min. Neglecting time costs, the combination of DarkNet19 and SGDM has proven to be the best ensuring the accuracy of 99.97%.